🔍 Read the full analysis: Why The Gemini 4 Carbon AI Model Score Doesn’t Tell The Whole Story on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Gemini 4 Carbon scored 59 on Artificial Analysis’s Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The index combines nine tests; Carbon leads four and trails on five, costs more than some rivals per index task and is not yet publicly available.
Gemini 4 Carbon has taken the top spot in Artificial Analysis’s Intelligence Index v5, scoring 59 points—one ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The ranking is a narrow lead across a particular set of benchmarks, not a finding that Carbon outperforms every rival at every task; Google is also limiting initial access to a small group of cyber defenders.
The index combines results from nine tests, grouped into knowledge work, coding, science and general capabilities, with each category carrying equal weight. Artificial Analysis says 60% of the version 5 evaluation uses private test sets or held-out answers. Many tasks involve extended work, such as using terminal tools, preparing reports or modifying code, rather than answering short prompts.
Carbon placed first on four of the nine tests: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. Its lead on the overall score is slim: Claude Opus 5.5 scored 58, GPT-6 Astra 57, GPT-6.1 Sol 55, and Google’s previous Gemini 4 Argon 50. On HLE-Diamond, Carbon scored 69% versus 68% for Opus 5.5; on Terminal-Bench-Science, it scored 65% versus 63% for GPT-6 Astra.
The reported cost per index task is $7.76 at standard prices. That is below Claude Opus 5.5’s $10.04, but above GPT-6 Astra’s $6.24 and GPT-6.1 Sol’s $1.27. Google’s introductory API price is listed as $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20 respectively. These are different measures: per-task index cost and token pricing should not be treated as interchangeable.
Infographic 01 · What the #1 is made of
Number one overall. First in four of nine tests.
The Intelligence Index is a decathlon score: nine tests, one number. Carbon wins on reasoning, science and careful knowledge, and trails on office work and coding.
The nine tests behind the score
Carbon at its highest setting (high) against the best model on each test.
| Test | Weight | Carbon | Leader or runner-up |
|---|---|---|---|
| HLE-Diamond Hard reasoning + knowledge questions | Weight 10% | Carbon 69% | Carbon leads Next: Claude Opus 5.5, 68% |
| Terminal-Bench-Science Science research work in a terminal | Weight 15% | Carbon 65% | Carbon leads Next: GPT-6 Astra, 63% |
| AA-Omniscience v1.1 Knowledge, wrong answers penalized | Weight 15% | Carbon 54 | Carbon leads Next: Claude Opus 5.5, 48 |
| AutomationBench-AA Business workflows across apps | Weight 5% | Carbon 78% | Carbon leads Next: Gemini 4 Argon, 78% |
| AA-Briefcase Office-style knowledge work | Weight 15% | Carbon 1,609 Elo | Behind Claude Sonnet 5.5, 1,812 |
| FrontierCode 2.0 Code a maintainer would merge | Weight 15% | Carbon 41% | Behind Claude Opus 5.5, 46% |
| CritPt v2 (hard) Research-level physics | Weight 10% | Carbon 63% | Behind GPT-6 Astra, 71% |
| Terminal-Bench 4.0 Agentic terminal tasks | Weight 10% | Carbon 65% | Behind Kimi K3.1 Pro, 67% |
| GDP.pdf Reasoning over documents | Weight 5% | Carbon 28% | Behind Muse Wonder Full, 33% |
The surprise in the knowledge test
- Accuracy: Claude Fable 5.1 68%, Claude Opus 5.5 66%, GPT-6.1 Sol and GPT-6 Astra 62%, Carbon 58%
- Hallucination rate: Carbon 11%, Claude Opus 5.5 and GPT-6 Astra about 50%
- Carbon wins because it rarely bluffs, not because it knows the most
Read the weights
- Each of the four blocks counts 25%: knowledge work, coding, science, general
- Carbon trails in two of the four heaviest tests (15% each): office-style work and real-world code
- Hallucination rate = wrong answers among all questions not answered correctly
Source: Artificial Analysis, 28 Oct 2026; Index v5 announcement, 27 Oct 2026; Artificial Analysis model data, retrieved 29 Oct 2026. Scores rounded as in the article; Carbon tested at high, other models at their highest tested setting.
Powered by Thorsten Meyer AI
A Narrow Lead, Not a Universal Win
The result matters because benchmark rankings can influence which models companies test, buy or build into products. But Carbon’s one-point margin does not establish a clear advantage across the work those organisations need to do. The index’s composite score rewards performance on its chosen tests and weighting; a business with different tasks, tools or cost limits may get a different result.
Carbon’s performance also involves trade-offs. It loses to Claude models on the index’s office-style knowledge-work test, trails Claude Opus 5.5 on real-world coding, and does not lead the physics-research, terminal-task or document-reasoning tests listed in the source. For purchasers, the practical question is less “Which model is first?” than how reliably and economically it completes their own tasks.
The gap between index cost and published token rates is relevant too. A task that uses many tokens or requires repeated steps can cost more than a simple prompt. The source’s figures offer a comparison under benchmark conditions, not a guarantee of what any individual customer will pay in production.
Products Worth Considering
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
Amazon Kindle Scribe Colorsoft 64GB (1st Generation)
DJI Osmo Pocket 3 Creator Combo Vlogging Camera with Mic 2, 1 CMOS, 4K/120fps, 3-Axis Stabilization, 128GB Card
As an affiliate, we earn on qualifying purchases.
How the Nine-Test Index Works
Artificial Analysis launched version 5 on October 27, rebuilding its leaderboard around nine evaluations divided into four equally weighted blocks. Knowledge work, coding, science and general capability each contribute 25% of the combined score. Within each block, individual tests have different weights, so the overall ranking reflects both how models perform and how the index allocates emphasis.
That design helps explain why one headline number can conceal uneven results. Carbon’s four first-place finishes coexist with five tests where another model performs better. Results also refer to a particular model setting: the source compares Carbon in a high-reasoning configuration with rivals at specified settings such as “max.” A change in model settings, test selection or weighting could change the ordering.
On AA-Omniscience, Carbon’s win is tied to how the evaluation treats uncertainty. The source reports Carbon answered 58% correctly, while several rival models had higher plain accuracy. But the test penalises wrong answers, and Carbon’s reported hallucination rate—defined there as wrong answers among questions not answered correctly—was 11%, compared with about 50% for Claude Opus 5.5 and GPT-6 Astra. The score therefore rewards cautious answers, not just the number of facts a model recalls.
“Choose by task and by cost per finished task, not by rank.”
— Thorsten Meyer, source article
Access and Real-World Costs Remain Open
Public availability is not yet dated. The source says Google plans to give Carbon first to “a subset of cyber defenders,” then to trusted partners, followed by developers, businesses and consumers. It gives no dates or details on how people qualify for those stages. Until access expands, most users cannot directly test whether the model suits their needs.
The material also does not provide enough detail to independently assess how closely benchmark performance predicts results in a customer’s specific setting. The cost-per-task figure is tied to the index, while actual bills depend on workload and token use. The available text ends partway through a comparison of coding-agent results, so it does not supply the complete explanation or all relevant conditions for that claim.
Finally, the reported score is a snapshot of the models and test version compared at publication. It does not show how rankings will change as models, tools or the benchmark itself are updated.
Google’s Staged Rollout Awaits Dates
The next practical milestone is Google’s access rollout, beginning with a limited group of cyber defenders and then moving to trusted partners and wider user groups, according to the source. Google has not supplied a schedule in the material provided. Later availability would let organisations compare Carbon against competitors on their own workflows rather than relying only on index results.
For companies considering a switch, the source recommends testing the model on representative work and tracking the cost per completed task, including accuracy and the amount of human correction required. Readers should also watch for more detail on availability, final API pricing and complete benchmark comparisons. For everyday users, there is no immediate change while public access remains unavailable.
Key Questions
What does Gemini 4 Carbon’s 59-point score mean?
It is Carbon’s combined result on Artificial Analysis’s nine-test Intelligence Index v5. It placed one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra on that index; it is not a score for every kind of AI task.
Does Carbon perform best on all nine tests?
No. The source says Carbon leads four of nine tests and trails other models on five, including office-style knowledge work and real-world coding.
Why can Carbon win the knowledge test despite lower accuracy?
AA-Omniscience penalises wrong answers. The source reports that Carbon’s lower hallucination rate helps its score, even though several competing models answered a higher share of questions correctly.
How much does Carbon cost?
The source lists $7.76 per index task at standard prices and introductory API rates of $2 per million input tokens and $10 per million output tokens, later listed as $4 and $20. A customer’s actual costs will depend on usage.
Can the public use Gemini 4 Carbon now?
Not according to the source material. Google is reported to be starting with a subset of cyber defenders and expanding access in stages, but no public release date is given.
Source: ThorstenMeyerAI.com
